[Excerpted from my book ‘Research Methods in Remote Sensing’]
Ontology
Ontology is the largest branch of metaphysics in philosophy, and traditionally deals with questions of existence or reality. Ontology as a branch of philosophy is the science of what is, of the kinds and structures of objects, properties, events, processes and relations in every area of reality. Ontology is often used by philosophers as a synonym for metaphysics―a term which was used by early students of Aristotle to refer to what Aristotle himself called first philosophy. Sometimes ontology is used in a broader sense, to refer to the study of what might exist, where metaphysics is used for the study of which of the various alternative possibilities is true of reality (Smith 1999). Ontology thus provides the basis for exchange of information, and is a fundamental pre-requisite to description and explanation, in science and elsewhere.
In simple words, ontology seeks the classification of entities. Typically, philosophical ontologists produce theories that are very much like scientific theories, but of a far more general nature. Ontology is both a branch of philosophy and a fast-growing component of computer science concerned with the development of formal representations of the entities and relations existing in a variety of application domains. Ontology has been shown to have considerable potential on the level of both pure research and applications. It provides foundations for diverse technologies in areas such as information integration, natural language processing, data annotation, and the construction of intelligent computer systems.
Recently, the term ontology has been used by information scientists to refer to canonical descriptions of knowledge domains, or associated classificatory theories. In this sense, ontology is “a neutral and computationally tractable description or theory of a given domain which can be accepted and reused by all information gatherers in that domain” (Smith 1999).
Often it is said that remote sensing can provide the ‘true’ representation of the earth’s surface. This statement is never true. Remote sensing provides an impression of the earth-surface features in pictorial format. Pictures are not the real truth; e.g., picture of a flower and the flower itself are not same. Therefore, ontology of remote sensing is primarily the inquiry about the existence or reality through the images. This inquiry is directed towards understanding and defining earth-surface features, spatial relations, processes, their categories, and so on. It would include not only the basic data models, concepts, and representations or classifications of earth-surface features, but also the ontological principles.
Objects and Fields
Remote sensing requires the classification (either object-based or field-based) of earth-surface features. The most widely accepted conceptual data model for spatial information considers that the geographic reality is represented as either fully definable entities (objects) or continuous spatial variations (fields). Objects are with discrete boundaries represented by geometric features; e.g., Roads, buildings, water bodies, etc. Fields are continuous phenomena such as elevation, temperature, and soil chemistry; they exist everywhere. These statements are very simple to the geospatial community. Unfortunately, remote sensing cannot handle them with that much of simplicity.
If we consider individual bands of remote sensing images, they are two-dimensional functions, arising from the sampled response of a region of the earth to an external energy source (the sun or a radar beam) as measured by a passive or active sensor, respectively. In case of thermal and passive microwave remote sensing, the earth itself is the source of energy. Whatever the source of energy or the techniques involved in capturing that energy, remote sensing images are always continuous, not discrete. Further the properties of each sensor (i.e., the number of bands as well as the spectral, temporal, radiometric, and spatial resolutions) are the results of a compromise between the needs of various research communities and the availability of sensor technology. The continuous variation of the spectral response of the land-cover, which is the specific phenomena captured by the pixel values, often misses what a domain scientist considers as relevant. These measurements are merely components of the more complex information content of an image. Most image classification techniques do not rely explicitly on the conversion between digital counts (pixel values) and the actual energy captured by the sensor, but they use the digital counts to extract features. As a consequence, viewing images as fields of values of reflected energy is insufficient for their ontological characterization (Câmara et al. 2001). The limitations of the field perspective to the ontology of images have led some researchers to view a remotely sensed image as a container of an implicit set of objects, which are extracted by manual or semi-automated analysis procedures. But important to realize, the real-world boundaries exist independent of human cognitive acts. Further, measurement of reflected/emitted light and the identification of objects via manual or semi-automated analysis are independent, not controlled or operated by one another. For example, identification of objects does not consider the real atmospheric contributions that might have recorded by the sensor. Furthermore, the pixel is a generalized representation of reality; mostly they are mixed in nature―mixel (mixed pixel). A pixel cannot be divided further to represent an object smaller than the pixel size.
Classes
Class refers to a group of features identical or similar types that have taxonomic significance. Classification is an abstraction mechanism that maps individuals into a common class. Classes are often interrelated by a generalization processes, capturing different levels of detail about the same individuals. For example, deciduous and evergreen are two classes that can be generalized as forest. In this case, ontology describes a hierarchy of concepts created by a generalization process. Although the object perspective captures a fundamental component of the ontology of images and forms a basis for a large set of image classification techniques, it is still incomplete. In many cases, there is no corresponding object in the world, since we deal with purely physical phenomena; e.g., vegetation health—that may change in every pixel. Actually, object-based classification is an attempt for geospatial segmentation of earth-surface features based on the pixels and their digital values. This helps us to understand the physical earth-surface as a distinctly segmented space. This concept overlooks the existence of ‘mixel’ (mixed pixel).
Field-based sub-pixel classification, on the other hand, considers the geographic space as continuous phenomena. It tries to quantify the amount of a feature or material within a pixel. The entire pixel or a part of it may be occupied by a material. Therefore, it calculates the percentage or amount of that material within a pixel. This concept handles the problem of mixel well, but ignores the identification of other surface materials. Therefore, there is no spatial segmentation among the features, and thus, no objects in existence.
The preceding discussion shows that neither of the classification approaches is sufficient by itself to support the full process of knowledge representation for remotely sensed images. The underlying reason is that images have a dual nature: earth surface features within them can be interpreted as fields as well as objects; although, they are fields at the measurement level.
Relations
Remote sensing ontologies are different from many of other ontologies in that it embraces spatial relations that play a major role in the geospatial domain. Earth-surface features can be connected or contiguous, scattered or separated, closed or open, near or far, and so on. These relations not only exist in geographic form but also in concepts. For example, two large forests may be connected by a narrow forested passage (geographic form), and these two forests may be connected in the sense that both are of same species (conceptual form). Therefore, ‘part and whole’ relations and other spatial relations are needed to be described in remote sensing ontologies.
Functions
Functions are mappings or transformations applied on remotely sensed images, and can be of many types. Basic image processing functions, such as filtering, principal component transformations, resampling, or interpolation, are familiar to and used by almost all researchers working with remote sensing data. Other generic functions exist in the statistical analyses of spatial domain, e.g., spatial statistics or spatial metrics. Of more significance here are functions that map the state of earth’s surface can be applied as a process because earth-surface features are not static―they are continuously changing with the change of time. For example, spatial metrics can determine the built-up pattern; but how the built-up pattern changes with time is essentially a process that requires temporal consideration. Our concept of remote sensing ontology thus includes notions both of form (shape/size/pattern) and process. The process does not refer only to historical process but simulations for the future as well.
Câmara et al. (2001) argued that “a geographic landscape is an ever-changing scenario, and the process of data capture by remote sensing satellites implies that an image is a measurement that captures snapshots of change trajectories. Therefore, the focus of the ontological characterization of images should be on the search for changes instead of the search for content. The emphasis of such ontologies should not be placed on simple object matching and identification procedures, but on capturing dynamics over a finite landscape.”
Image Ontology
Câmara et al. (2001) proposed a multi-level ontology for images, based on the concept of action-driven ontologies for GIS (Câmara et al. 2000). The authors considered that remote sensing images are ontological instruments to capture landscape dynamics. The proposal takes into account that images have a particular, distinct description independent of the domain ontology a scientist would employ to extract information. The ontology domain for images has three interrelated components:
(1) Physical ontology – describes the physical process of the image creation, focusing the knowledge about the relation between the reflected energy by terrain surface and measures obtained by the sensor.
(2) Structural Ontology – contemplates geometric, functional and descriptive structures that can be extracted using techniques for feature extraction, segmentation, classification, and so on.
(3) Method Ontology – it is composed of a set of algorithms (that perform transformations from the physical level to the structural level) and data structures that represent reusable knowledge in the form of image processing techniques (filtering, smoothing, and others).
The algorithms that are part of the method ontology, perform transformations from the physical level to the structural level, a process than can be called structural identification. When applied to an image (or a set of images), this process results in a set of structures strongly related to the measurement device properties and its interaction with the physical landscape. These structures may be geometric (e.g., regions extracted by a segmentation procedure, i.e., per-pixel classification) or functional (e.g., normalized differential vegetation index or sub-pixel classification) (Câmara et al. 2001).
Image Mining and Image Ontology
Image mining deals with extraction of implicit knowledge, image data relationship or other patterns not explicitly stored in images and uses ideas from computer vision, image processing, image retrieval, data mining, machine learning, databases and artificial intelligence. The fundamental challenge in image mining is to determine how low-level pixel representation contained in an image or an image sequence can be effectively and efficiently processed to identify high-level spatial objects and relationships. Typical image mining process involves pre-processing, transformations and feature extraction, mining (to discover significant patterns out of extracted features), evaluation and interpretation and obtaining the final knowledge. Various techniques from existing domains are also applied to image mining and include object recognition, learning, clustering and classification, just to name a few (Zhang et al. 2002).
Extracting information from images remains a complex and tedious process; sometimes inferior in respect of our needs. Our capacity to build sophisticated remote sensing sensors is not matched by our means of producing information from these data sources. Currently, most image processing techniques are designed to operate on a single image, and we have few algorithms and techniques for handling multi-temporal images. This situation has lead to a ‘knowledge gap’ in the process of deriving information from images and digital maps (MacDonald 2002). This ‘knowledge gap’ has arisen because there are currently few techniques for image mining and information extraction in large image datasets; thus we are failing to exploit our large remote sensing archives. Image ontology facilitates the deployment of the concept for various classes of models for the information extraction from the remote sensing imagery. Silva and Câmara (2004) presented the architecture of ontology based image mining. Durand et al. (2007) also explained the ontology-based object recognition for remote sensing image interpretation.
REFERENCES
Câmara, G., Egenhofer, M., Fonseca, F., and Monteiro, A.M.V. (2001). What’s in an Image?. In Spatial Information Theory: Foundations of Geographic Information Science, International Conference, COSIT.
Câmara, G.M., Monteiro, A.M.V., Paiva, J.A.C, and Souza, R.C.M. (2000, October). Action-Driven Ontologies of the Geographical Space: Beyond the Field-Object Debate. In GIScience, Savanah, GA: AAG, pp. 52–54.
Durand, N., Derivaux, S., Forestier, G., Wemmert, C., & Gancarski, P. O. Boussaid D, et A. Puissant (2007). Ontology-based Object Recognition for Remote Sensing Image Interpretation. In IEEE International Conference on Tools with Artificial Intelligence, Patras, Greece, pp. 472–479.
MacDonald, J. (2002). The Earth Observation Business and the Forces that Impact it. Earth Observation Business Network.
Silva, M.P.S., and Câmara, G. (2004). Remote Sensing Image Mining Using Ontologies. Online. URL:http://www.dpi.inpe.br/~mpss/artigos/Im ... g2004.pdf.
Smith, B., (1999). Ontology: Philosophical and Computational. Unpublished manuscript. URL:http://wings.buffalo.edu/philosophy/fac ... ogies.htm.
Zhang, J., Hsu, W., and Lee, M. (2002). Image Mining: Trends and Developments. Dordrecht: Kluwer Academic.
Ontology
Ontology is the largest branch of metaphysics in philosophy, and traditionally deals with questions of existence or reality. Ontology as a branch of philosophy is the science of what is, of the kinds and structures of objects, properties, events, processes and relations in every area of reality. Ontology is often used by philosophers as a synonym for metaphysics―a term which was used by early students of Aristotle to refer to what Aristotle himself called first philosophy. Sometimes ontology is used in a broader sense, to refer to the study of what might exist, where metaphysics is used for the study of which of the various alternative possibilities is true of reality (Smith 1999). Ontology thus provides the basis for exchange of information, and is a fundamental pre-requisite to description and explanation, in science and elsewhere.
In simple words, ontology seeks the classification of entities. Typically, philosophical ontologists produce theories that are very much like scientific theories, but of a far more general nature. Ontology is both a branch of philosophy and a fast-growing component of computer science concerned with the development of formal representations of the entities and relations existing in a variety of application domains. Ontology has been shown to have considerable potential on the level of both pure research and applications. It provides foundations for diverse technologies in areas such as information integration, natural language processing, data annotation, and the construction of intelligent computer systems.
Recently, the term ontology has been used by information scientists to refer to canonical descriptions of knowledge domains, or associated classificatory theories. In this sense, ontology is “a neutral and computationally tractable description or theory of a given domain which can be accepted and reused by all information gatherers in that domain” (Smith 1999).
Often it is said that remote sensing can provide the ‘true’ representation of the earth’s surface. This statement is never true. Remote sensing provides an impression of the earth-surface features in pictorial format. Pictures are not the real truth; e.g., picture of a flower and the flower itself are not same. Therefore, ontology of remote sensing is primarily the inquiry about the existence or reality through the images. This inquiry is directed towards understanding and defining earth-surface features, spatial relations, processes, their categories, and so on. It would include not only the basic data models, concepts, and representations or classifications of earth-surface features, but also the ontological principles.
Objects and Fields
Remote sensing requires the classification (either object-based or field-based) of earth-surface features. The most widely accepted conceptual data model for spatial information considers that the geographic reality is represented as either fully definable entities (objects) or continuous spatial variations (fields). Objects are with discrete boundaries represented by geometric features; e.g., Roads, buildings, water bodies, etc. Fields are continuous phenomena such as elevation, temperature, and soil chemistry; they exist everywhere. These statements are very simple to the geospatial community. Unfortunately, remote sensing cannot handle them with that much of simplicity.
If we consider individual bands of remote sensing images, they are two-dimensional functions, arising from the sampled response of a region of the earth to an external energy source (the sun or a radar beam) as measured by a passive or active sensor, respectively. In case of thermal and passive microwave remote sensing, the earth itself is the source of energy. Whatever the source of energy or the techniques involved in capturing that energy, remote sensing images are always continuous, not discrete. Further the properties of each sensor (i.e., the number of bands as well as the spectral, temporal, radiometric, and spatial resolutions) are the results of a compromise between the needs of various research communities and the availability of sensor technology. The continuous variation of the spectral response of the land-cover, which is the specific phenomena captured by the pixel values, often misses what a domain scientist considers as relevant. These measurements are merely components of the more complex information content of an image. Most image classification techniques do not rely explicitly on the conversion between digital counts (pixel values) and the actual energy captured by the sensor, but they use the digital counts to extract features. As a consequence, viewing images as fields of values of reflected energy is insufficient for their ontological characterization (Câmara et al. 2001). The limitations of the field perspective to the ontology of images have led some researchers to view a remotely sensed image as a container of an implicit set of objects, which are extracted by manual or semi-automated analysis procedures. But important to realize, the real-world boundaries exist independent of human cognitive acts. Further, measurement of reflected/emitted light and the identification of objects via manual or semi-automated analysis are independent, not controlled or operated by one another. For example, identification of objects does not consider the real atmospheric contributions that might have recorded by the sensor. Furthermore, the pixel is a generalized representation of reality; mostly they are mixed in nature―mixel (mixed pixel). A pixel cannot be divided further to represent an object smaller than the pixel size.
Classes
Class refers to a group of features identical or similar types that have taxonomic significance. Classification is an abstraction mechanism that maps individuals into a common class. Classes are often interrelated by a generalization processes, capturing different levels of detail about the same individuals. For example, deciduous and evergreen are two classes that can be generalized as forest. In this case, ontology describes a hierarchy of concepts created by a generalization process. Although the object perspective captures a fundamental component of the ontology of images and forms a basis for a large set of image classification techniques, it is still incomplete. In many cases, there is no corresponding object in the world, since we deal with purely physical phenomena; e.g., vegetation health—that may change in every pixel. Actually, object-based classification is an attempt for geospatial segmentation of earth-surface features based on the pixels and their digital values. This helps us to understand the physical earth-surface as a distinctly segmented space. This concept overlooks the existence of ‘mixel’ (mixed pixel).
Field-based sub-pixel classification, on the other hand, considers the geographic space as continuous phenomena. It tries to quantify the amount of a feature or material within a pixel. The entire pixel or a part of it may be occupied by a material. Therefore, it calculates the percentage or amount of that material within a pixel. This concept handles the problem of mixel well, but ignores the identification of other surface materials. Therefore, there is no spatial segmentation among the features, and thus, no objects in existence.
The preceding discussion shows that neither of the classification approaches is sufficient by itself to support the full process of knowledge representation for remotely sensed images. The underlying reason is that images have a dual nature: earth surface features within them can be interpreted as fields as well as objects; although, they are fields at the measurement level.
Relations
Remote sensing ontologies are different from many of other ontologies in that it embraces spatial relations that play a major role in the geospatial domain. Earth-surface features can be connected or contiguous, scattered or separated, closed or open, near or far, and so on. These relations not only exist in geographic form but also in concepts. For example, two large forests may be connected by a narrow forested passage (geographic form), and these two forests may be connected in the sense that both are of same species (conceptual form). Therefore, ‘part and whole’ relations and other spatial relations are needed to be described in remote sensing ontologies.
Functions
Functions are mappings or transformations applied on remotely sensed images, and can be of many types. Basic image processing functions, such as filtering, principal component transformations, resampling, or interpolation, are familiar to and used by almost all researchers working with remote sensing data. Other generic functions exist in the statistical analyses of spatial domain, e.g., spatial statistics or spatial metrics. Of more significance here are functions that map the state of earth’s surface can be applied as a process because earth-surface features are not static―they are continuously changing with the change of time. For example, spatial metrics can determine the built-up pattern; but how the built-up pattern changes with time is essentially a process that requires temporal consideration. Our concept of remote sensing ontology thus includes notions both of form (shape/size/pattern) and process. The process does not refer only to historical process but simulations for the future as well.
Câmara et al. (2001) argued that “a geographic landscape is an ever-changing scenario, and the process of data capture by remote sensing satellites implies that an image is a measurement that captures snapshots of change trajectories. Therefore, the focus of the ontological characterization of images should be on the search for changes instead of the search for content. The emphasis of such ontologies should not be placed on simple object matching and identification procedures, but on capturing dynamics over a finite landscape.”
Image Ontology
Câmara et al. (2001) proposed a multi-level ontology for images, based on the concept of action-driven ontologies for GIS (Câmara et al. 2000). The authors considered that remote sensing images are ontological instruments to capture landscape dynamics. The proposal takes into account that images have a particular, distinct description independent of the domain ontology a scientist would employ to extract information. The ontology domain for images has three interrelated components:
(1) Physical ontology – describes the physical process of the image creation, focusing the knowledge about the relation between the reflected energy by terrain surface and measures obtained by the sensor.
(2) Structural Ontology – contemplates geometric, functional and descriptive structures that can be extracted using techniques for feature extraction, segmentation, classification, and so on.
(3) Method Ontology – it is composed of a set of algorithms (that perform transformations from the physical level to the structural level) and data structures that represent reusable knowledge in the form of image processing techniques (filtering, smoothing, and others).
The algorithms that are part of the method ontology, perform transformations from the physical level to the structural level, a process than can be called structural identification. When applied to an image (or a set of images), this process results in a set of structures strongly related to the measurement device properties and its interaction with the physical landscape. These structures may be geometric (e.g., regions extracted by a segmentation procedure, i.e., per-pixel classification) or functional (e.g., normalized differential vegetation index or sub-pixel classification) (Câmara et al. 2001).
Image Mining and Image Ontology
Image mining deals with extraction of implicit knowledge, image data relationship or other patterns not explicitly stored in images and uses ideas from computer vision, image processing, image retrieval, data mining, machine learning, databases and artificial intelligence. The fundamental challenge in image mining is to determine how low-level pixel representation contained in an image or an image sequence can be effectively and efficiently processed to identify high-level spatial objects and relationships. Typical image mining process involves pre-processing, transformations and feature extraction, mining (to discover significant patterns out of extracted features), evaluation and interpretation and obtaining the final knowledge. Various techniques from existing domains are also applied to image mining and include object recognition, learning, clustering and classification, just to name a few (Zhang et al. 2002).
Extracting information from images remains a complex and tedious process; sometimes inferior in respect of our needs. Our capacity to build sophisticated remote sensing sensors is not matched by our means of producing information from these data sources. Currently, most image processing techniques are designed to operate on a single image, and we have few algorithms and techniques for handling multi-temporal images. This situation has lead to a ‘knowledge gap’ in the process of deriving information from images and digital maps (MacDonald 2002). This ‘knowledge gap’ has arisen because there are currently few techniques for image mining and information extraction in large image datasets; thus we are failing to exploit our large remote sensing archives. Image ontology facilitates the deployment of the concept for various classes of models for the information extraction from the remote sensing imagery. Silva and Câmara (2004) presented the architecture of ontology based image mining. Durand et al. (2007) also explained the ontology-based object recognition for remote sensing image interpretation.
REFERENCES
Câmara, G., Egenhofer, M., Fonseca, F., and Monteiro, A.M.V. (2001). What’s in an Image?. In Spatial Information Theory: Foundations of Geographic Information Science, International Conference, COSIT.
Câmara, G.M., Monteiro, A.M.V., Paiva, J.A.C, and Souza, R.C.M. (2000, October). Action-Driven Ontologies of the Geographical Space: Beyond the Field-Object Debate. In GIScience, Savanah, GA: AAG, pp. 52–54.
Durand, N., Derivaux, S., Forestier, G., Wemmert, C., & Gancarski, P. O. Boussaid D, et A. Puissant (2007). Ontology-based Object Recognition for Remote Sensing Image Interpretation. In IEEE International Conference on Tools with Artificial Intelligence, Patras, Greece, pp. 472–479.
MacDonald, J. (2002). The Earth Observation Business and the Forces that Impact it. Earth Observation Business Network.
Silva, M.P.S., and Câmara, G. (2004). Remote Sensing Image Mining Using Ontologies. Online. URL:http://www.dpi.inpe.br/~mpss/artigos/Im ... g2004.pdf.
Smith, B., (1999). Ontology: Philosophical and Computational. Unpublished manuscript. URL:http://wings.buffalo.edu/philosophy/fac ... ogies.htm.
Zhang, J., Hsu, W., and Lee, M. (2002). Image Mining: Trends and Developments. Dordrecht: Kluwer Academic.